Remote sensing image small target detection method based on high-dimensional space and multi-scale fusion

Abstract

The invention discloses a remote sensing image small target detection method based on high-dimensional space and multi-scale fusion, and the method comprises the steps: obtaining a certain number of aviation and satellite remote sensing images in different scenes to construct a data set, and carrying out the division according to a preset proportion; a high-dimensional space information and multi-scale feature fusion model is constructed, and in a feature fusion network of the model, a Sobel detection operator and dynamic convolution are adopted, and a C3k2-EFEM module is designed; an MIF module used for fusing multi-dimensional information is designed by using a full-field observer; a dynamic up-sampling DUB module is designed; and training, testing and verifying the model, and applying the tested model to remote sensing image small target detection. Through the effective use of the three modules, the detection effect of the small target model is enhanced, the network can stably cope with the detection scene of the small target, and the method can give consideration to the detection performance and the detection processing speed in the small target detection of the remote sensing image.

Description

Technical Field

[0001] This invention relates to the field of remote sensing target recognition technology, and in particular to a method for detecting small targets in remote sensing images based on high-dimensional space and multi-scale fusion. Background Technology

[0002] Advances in remote sensing technology have significantly increased the volume and fidelity of image data used for representing Earth's surface features, making object detection technology based on remote sensing images a prominent research area. The rise of remote sensing technology has propelled its application across numerous disciplines, including environmental and disaster monitoring, intelligent transportation, and military reconnaissance. Therefore, remote sensing images have become an indispensable component of contemporary geospatial analysis. However, the transformation of remote sensing image detection into practical applications still faces many key obstacles. Targets in these images (including but not limited to ships, aircraft, and vehicles) generally exhibit small target characteristics, often occupying a negligible proportion of image pixels, making them highly susceptible to losing crucial information during convolutional downsampling. Furthermore, remote sensing images often depict complex natural environments; the visual similarity between a specific target and its surrounding terrain can trigger background noise interference or target overlap, leading to false positives or false negatives. Simultaneously, the diversity of target scales and the complexity of their types demand that detection algorithms possess stronger generalization capabilities. Therefore, maintaining a reasonable balance between detection accuracy and processing speed remains a core challenge in the field of remote sensing target recognition. In the field of remote sensing image detection, early methods mainly relied on manual feature extraction and shallow machine learning classifiers, widely employing methods such as Scale Invariant Feature Transform (SIFT) and Histogram of Oriented Gradient (HOG) for detection. However, traditional detection techniques have limitations when facing complex environments and diverse targets.

[0003] Existing deep learning-based object detection algorithms are gradually replacing traditional detection methods. Compared to traditional methods, these deep learning methods utilize multi-layer network structures and large datasets for autonomous learning and feature extraction. These algorithms first identify regions where targets may appear, and then perform target detection based on region proposals. While their detection accuracy is high, their detection speed is relatively slow, and they perform poorly in detecting small targets. A common solution is to add an additional detection head at the top of FPN / PANet (i.e., the shallowest, highest-resolution feature map). Since small targets are easily lost during downsampling, using higher-resolution feature maps can preserve more detailed information about small targets. Although this method directly addresses the problem of information loss in small targets and provides the most direct and detailed features for detecting small targets in remote sensing, the high-resolution feature maps in remote sensing images contain a large number of pixels, significantly increasing the computational load and parameter count, making them unsuitable for deployment on mobile devices. They also tend to have more noise in shallow features, potentially introducing more false positives. Therefore, existing deep learning detection methods cannot achieve a balance between detection accuracy and efficiency for small targets in remote sensing image scenarios.

[0004] In the task of small target detection in remote sensing images, the small pixel ratio of targets and the limited available visual features are key factors restricting detection performance. Relying solely on two-dimensional image data often fails to achieve accurate identification. While three-dimensional data containing depth information can provide a more complete geometric structure and spatial context, significantly improving positioning accuracy and scene understanding, its high computational complexity and stringent hardware requirements make balancing accuracy and efficiency a core issue in practical applications. The C3k2 module in the YOLO11 network optimizes the convolutional layer structure and gradient flow path, enhancing the preservation of spatial details while maintaining a relatively light computational load, making it particularly suitable for small target detection scenarios. However, when faced with the common problems of dense targets, cluttered backgrounds, or occlusion and overlap between targets in remote sensing images, the detection performance of this module still shows a significant decline. Furthermore, the influence of cloud and fog interference and sensor noise in remote sensing images further exacerbates the feature blurring of dense small targets, leading to the risk of false positives and false negatives. Existing methods often attempt to recover degraded areas through image reconstruction techniques, but traditional reconstruction methods are difficult to meet real-time processing requirements.

[0005] Therefore, exploring multi-scale fusion and feature enhancement mechanisms that can effectively enhance the representation of small target features in complex remote sensing environments while balancing computational efficiency and detection accuracy has become one of the important research directions for improving the practicality of detection models, and further exploration is needed. Summary of the Invention

[0006] To address the aforementioned problems and technical requirements, the inventors have proposed a method for detecting small targets in remote sensing images based on high-dimensional space and multi-scale fusion. This method improves the YOLO11 network to solve the problem that remote sensing images generally have significant spatial resolution differences and a large number of compact targets, which makes the detection task complex, especially the detection of small targets.

[0007] The technical solution of the present invention is as follows: A method for detecting small targets in remote sensing images based on high-dimensional space and multi-scale fusion, specifically including the following steps: Step 1: Acquire a certain number of aerial and satellite remote sensing images of different scenes, scale all images to a preset size (e.g., 640*640), and normalize the pixel values ​​to form a dataset; Step 2: Divide the dataset into training, validation, and test sets according to a predetermined ratio; Step 3: Construct a MIFF-YOLO model that fuses high-dimensional spatial information with multi-scale features. This model includes a sequentially connected feature extraction network (backbone network), a feature fusion network, and three decoupled detection heads; wherein: The feature extraction network incorporates a C2PSA module, which enhances feature representation capabilities through a position-sensitive attention mechanism. Furthermore, the module employs a multi-head attention structure combined with a feedforward neural network, enabling the model to focus on key information at different spatial locations. In addition, the feedforward neural network maps features to a high-dimensional space and mines complex relationships between features through nonlinear transformations, thereby constructing a more discriminative feature representation.

[0008] The feature fusion network employs a MIFF (Multi-Module Fusion) network, which integrates multiple modules to fully fuse the feature information output by the feature extraction network. Specifically, the original C3k2 module in the feature fusion network is replaced with an improved C3k2-EFEM module, which uses the Sobel detection operator and dynamic convolution ODConv to effectively improve the network's detection performance. A MIF module is designed using a full-domain observer FDT (Focus-Driven Observer) to enable cross-dimensional fusion of feature information, enhancing the network's feature extraction capability to a certain extent. The MIF module ultimately uses 3DMaxPool to reduce computational cost while maintaining feature expressiveness. A dynamic upsampling DUB (Dynamic Upsampling Unit) module is designed using the Dysample operator. The Dysample module upsamples the feature map's dimensions and further enhances the upsampled feature map using a 3x3 depthwise convolution (DwConv), thereby improving the network's information flow and expressive power.

[0009] In a decoupled detection head architecture, the three detection heads typically utilize multiple parallel branches to process feature information at different scales. This architecture separates the tasks of target localization and category prediction, allowing them to be learned by independent network modules, and finally fusing the prediction results of both at the output stage. This task decoupling design not only significantly reduces the computational burden and parameter size of the model, but also helps improve the model's adaptability to different scenarios and its overall robustness.

[0010] Step 4: Input the training set into the MIFF-YOLO model for training, including: 1) Randomly initialize the network parameters of the model, and then input the training set images into the MIFF-YOLO model constructed in step 3 in batches; 2) Feature information is extracted by the feature extraction network, passes through the SPPF layer and the C2PSA attention layer, and then enters the Neck layer of the feature fusion network. Information from designated feature layers (P3, P4, P5) undergoes regular forward propagation and is also simultaneously fed into the MIF module for effective fusion with subsequent feature information. To enhance the feature fusion capability of the detection layer, an integrated attention mechanism module, FDT, is introduced into the MIF module. Simultaneously, the feature information undergoes upsampling processing by the DUB module before entering the C3k2-EFEM module.

[0011] 3) After co-optimizing the bounding box regression and classification tasks using the DIOU loss function and DFL loss, a task-aligned sample matching strategy is further employed to assign positive and negative samples. This strategy first calculates the intersection-union ratio (IU) between each predicted bounding box and its corresponding ground truth bounding box, and multiplies it by the classification confidence of the corresponding sample to obtain a task alignment score reflecting the consistency between classification and localization. Based on this, for each ground truth bounding box, the top N predicted samples with the highest alignment scores are selected as positive samples, and the rest are considered negative samples.

[0012] 4) During model training, the loss value is calculated using the aforementioned loss function. Based on this, the weight parameters of the unfrozen portions of the network are updated. To improve training efficiency, this method employs a cosine annealing strategy to dynamically adjust the learning rate. This mechanism uses the total number of training iterations as half a cycle, causing the learning rate to decrease smoothly along a cosine curve. This periodic adjustment not only helps the model converge stably but also effectively reduces the risk of getting trapped in local optima. Simultaneously, the continuous and gradual change in the learning rate provides a stable update environment for parameter optimization, further promoting the network's ability to learn feature representations.

[0013] 5) Repeat steps 1 to 4 until the set number of iterations is reached.

[0014] Step 5: The trained MIFF-YOLO model is tested using test and validation sets to obtain performance metrics and verify its effectiveness. The tested model is then applied to small target detection in remote sensing images. Specifically, during model training, the training data is trained following the aforementioned training process. After training, a reserved test set is used to systematically validate the model's performance. By quantitatively evaluating network models that meet preset performance thresholds, multiple core metrics, including average precision, recall, and detection speed, are obtained to comprehensively measure the model's overall performance in real-world application scenarios. Ultimately, this multi-scale feature fusion network enables efficient utilization of the high-dimensional spatial information of targets in remote sensing images, achieving effective detection of small targets in remote sensing images.

[0015] The further technical solution involves using FDT in the MIF module to simultaneously capture three types of detailed features: local, regional, and global. Its core lies in the combination of Regional Self-Attention (RSA) and Global Self-Attention (GSA). In RSA, the input feature map is first divided into uniformly sized, non-overlapping sub-windows. Where i ranges from 1 to M, and each sub-window covers an N×N spatial region. Within each sub-window, pointwise convolution and depthwise convolution are first used to generate the query matrix. Key matrix Sum matrix The dimensions of these matrices are all It can be expressed as the following formula:

[0016] in, Represents the query matrix. Represents the key matrix. Represents a value matrix, This indicates the reshaping operator. Represents a deep convolutional layer. This represents a point convolutional layer.

[0017] Based on the above, the region self-attention representation of each sub-window feature block is as follows:

[0018] in, express Activation function This represents a learnable parameter used to control the smoothness of the attention distribution.

[0019] A further technical solution is to introduce a sliding window strategy in RSA in order to break the boundary effect between sub-windows. That is, in the process of adjacent levels or iterations, the window is shifted by a certain offset to realize cross-window information flow and feature interaction.

[0020] A further technical solution involves using a global self-attention mechanism to perform multi-head attention computation simultaneously across the entire feature map, thereby obtaining macroscopic structural information and the dependencies between channels. In GSA, for the normalized features of the input layer... First, a query matrix is ​​generated through 1×1 pointwise convolution and then 3×3 depthwise convolution. Key matrix Sum matrix The shape was adjusted to have h heads, among which This represents the number of channels per head. The attention map is expanded along the channel direction, and then the vector is calculated. and The dot product is obtained The global attention graph; the entire process of global self-attention is represented as follows:

[0021] in, With the above Similarly, both represent a learnable parameter.

[0022] Its further technical solution is that the core of the DUB module is the dynamic upsampling operator Dysample, and its dynamic upsampling process includes: The input feature map is first passed through a 1×1 convolution. To generate offset Subsequently, the offset Used to adjust the standard sampling grid This yields a new dynamic sampling coordinate set. Using a standard sampling grid According to the sampling set Resampling is performed from the original low-resolution input feature map to generate the final high-resolution output feature map. The overall process can be represented by the following formula:

[0023]

[0024]

[0025] in, This represents the low-resolution input feature map to be upsampled. It is a 1×1 convolution. It is a sampling grid function. This method can effectively preserve the key details of small objects after interpolation and avoid image blurring caused by nearest-neighbor interpolation.

[0026] Its further technical solution is that the core of the C3k2-EFEM module is its internal EFEM (Efficient Front-End Module). EFEM enriches feature information through a dual-branch structure, namely the SobelConv branch and the MaxPool branch. The SobelConv branch uses the horizontal and vertical convolution kernels of the Sobel operator to explicitly extract edge features in the image; this process involves virtual dimensionality expansion of the feature map (…). T After performing convolutions in the X and Y directions respectively, the virtual dimension is removed. T' The operation ultimately merges the edge responses from both directions. The MaxPool branch extracts and preserves rich spatial detail information through an additional max-pooling layer.

[0027] The beneficial technical effects of this invention are: 1) To improve the detection accuracy of small targets, this method designs a MIF module for the neck region. This module aims to enhance the efficiency of multi-dimensional spatial feature information fusion while minimizing the impact on the number of parameters and computational cost.

[0028] 2) To overcome the problem of lost feature details of small targets during detection, this method replaces the original C3k2 module in the neck area with the C3k2-EFEM module. This module significantly improves the detection accuracy and efficiency of small targets while maintaining rich spatial details, thereby enhancing feature extraction and information fusion capabilities.

[0029] 3) To address the blurring issue caused by messy edge textures or complex features in images, this method replaces the original upsampling module with an efficient DUB module. While preserving the key details of small target features after interpolation, it enhances the model's understanding of the global context, thereby improving detection accuracy.

[0030] 4) Based on the integration of the above modules, this method reconstructs the feature fusion network architecture and designs a novel multi-dimensional information fusion neck structure. This neck structure adopts a multi-branch parallel architecture, achieving multi-scale information aggregation through cross-level feature interaction, significantly enhancing the model's perception and recognition performance of small targets. This design effectively solves the problem of easy loss of detailed features in traditional single-path systems, while improving the coherence of semantic representations between features of different scales. Attached Figure Description

[0031] Figure 1 This is a flowchart of the remote sensing image small target detection method proposed in this application; Figure 2This is a diagram of the MIFF-YOLO network structure for high-dimensional spatial information and multi-scale fusion proposed in this application; Figure 3 These are the overall and detailed diagrams of the MIF module proposed in this application, as well as the detailed diagram of the full-dimensional observer FDT; where: (1) is the feature fusion MIFF network diagram, and (2) is the detailed diagram of the full-dimensional observer FDT; Figure 4 This is a structural diagram of the improved C3k2-EFEM proposed in this application; Figure 5 This is the structure diagram of the efficient front-end module EFEM proposed in this application; Figure 6 This is a schematic diagram of the Sobel detection operator mechanism proposed in this application; Figure 7 This is a diagram of the dynamic convolution ODConv structure proposed in this application; Figure 8 This is a schematic diagram of the dynamic attention mechanism involved in ODConv proposed in this application; Figure 9 This is a structural diagram of the high-efficiency upsampling module DUB proposed in this application; Figure 10 This is a schematic diagram of Dysample, a core module involved in the DUB module.

[0032] Figure 11 The diagram shows the principle of the bilinear initialization strategy involved in the Dysample module; where: (a) is the starting point of each target position; (b) is a schematic diagram of the initial sampling position and offset range; and (c) is a schematic diagram of the offset range after local constraints. Detailed Implementation

[0033] The specific embodiments of the present invention will be further described below with reference to the accompanying drawings.

[0034] The presence of significant spatial resolution differences and numerous compact small targets in remote sensing images complicates the detection task. Existing target detection algorithms, particularly benchmark networks, suffer from insufficient ability to extract features from small targets. This embodiment effectively improves the performance of the C3k2 module in YOLO11 by employing the Sobel detection operator and dynamic convolution ODConv (Omni-dimensional Dynamic Convolution), designing the C3k2-EFEM (C3k2-Efficient Front-End Module). This allows the network to obtain important edge information and features while preserving rich spatial details.

[0035] Meanwhile, to address the lack of network feature fusion capabilities, the feature fusion module was reconstructed. To solve the problem of small targets being affected by complex environmental factors, a multidimensional information fusion module (MIF) was designed to enhance the network's ability to process multidimensional information across dimensions, thereby effectively reducing the interference of complex geospatial environments on the detection of small targets.

[0036] In addition, to address the problem of losing key details after multi-scale target magnification during remote sensing image detection, a Dynamic Upconvolution Block (DUB) was designed to enhance the network's information extraction capabilities during multi-scale target magnification, thereby improving the model's robustness and the efficiency of the upsampling process.

[0037] Based on the above improvements, a high-dimensional spatial information and multi-scale fusion network model, MIFF-YOLO (Multidimensional Information Feature Fusion - You Only Look Once), was constructed. First, the C3k2 module in the network feature fusion structure was improved by embedding an EFEM structure, enhancing the feature fusion capability of the original C3k2 while simultaneously reducing the model's parameter count and computational cost. Second, a feature fusion module, MIF, was constructed, embedding a Full Domain Transformer (FDT) observer to achieve multi-dimensional sequence information fusion, improving detection accuracy while minimizing model parameter count and computational cost. Finally, a dynamic upsampling DUB module was designed, dividing the input features into multiple groups along the channel dimension, with each group generating an independent offset to adapt to the differentiated needs of different semantic regions. Finally, based on the dynamically offset-adjusted coordinate grid, bilinear interpolation sampling was performed on the original features to generate high-resolution output, effectively solving the problem of losing key details after upsampling of multi-scale targets and improving the detection performance of small targets in extreme environments.

[0038] By effectively utilizing the three modules described above, the detection performance for small targets is enhanced, while also enabling the network to stably handle small target detection scenarios. This method balances detection performance and processing speed in remote sensing images, providing convenience for deployment on mobile devices.

[0039] Based on the above principles, this application provides a method for detecting small targets in remote sensing images based on high-dimensional space and multi-scale fusion. The detection process of this method is as follows: Figure 1 As shown, the specific implementation includes the following: like Figure 2The diagram shows the MIFF-YOLO network architecture. The MIFF-YOLO network mainly consists of a feature extraction network (backbone network), a feature fusion network, and three decoupled detection heads. The feature extraction network uses the CBS module, C3k2 module, SPPF module, and C2PSA module. The CBS module adds a batch normalization layer and a SilU activation function after each convolutional layer. The C3k2 module is composed of the CBS module and the Bottleneck module in a multi-branch structure.

[0040] The C3k2 module's Bottleneck in the feature fusion network is replaced with an EFEM structure, resulting in the C3k2-EFEM module, whose structure is as follows: Figure 4 As shown, the input features first undergo preliminary feature transformation through a first CBS module consisting of convolutional layers, batch normalization layers, and the SiLU activation function. Then, a channel splitting operation (Split) divides the feature map into two branches. One branch performs deep feature extraction through several cascaded EFEM modules, while the other retains the original split features. Finally, the deeply processed features are concatenated with the features from the skip connection branch along the channel dimension, and feature fusion and output are completed through a second CBS module. This design, through skip connections and a modular structure, achieves full fusion of multi-level features while maintaining effective gradient propagation.

[0041] The structure of the EFEM module is as follows: Figure 5 As shown, this module introduces the edge sampling operator Sobel, resulting in the SobelConv branch, which can explicitly extract edge features from the image. The specific process of SobelConv is as follows: Figure 6 As shown. Besides edge information, spatial information in the image is equally important. The EFEM module introduces an additional max-pooling branch, MaxPool, to extract spatial information. Unlike the SobelConv branch mentioned above, the convolution branch extracts features from the original image, preserving rich spatial details, and is especially suitable for small target detection in remote sensing images.

[0042] After fusing edge features and pooling features using the above operations, dynamic convolution ODConv is finally used to optimize the features. The principle is as follows: Figure 7 and Figure 8 As shown, ODConv employs a multi-dimensional dynamic attention mechanism, which not only dynamically adjusts the number of convolutional kernels but also involves three other dimensions of the kernels: spatial size, number of input channels, and number of output channels. This means that ODConv can more precisely adapt to the features of the input data, thereby improving the feature extraction effect.

[0043] The MIF module was also used in the neck area, and the FDT module was used for improvement. The principle is as follows: Figure 3 As shown in (1) and (2) in the figure. The MIF module can perform cross-dimensional processing on multi-dimensional information from the P3, P4 and P5 detection layers, and can obtain multi-scale features. At the same time, it uses 3D convolution to model the relationship between layers in the depth dimension to enhance feature interaction. Finally, 3DMaxPool is used to compress the computation while maintaining the feature representation capability. In addition, in order to improve the effect of the MIF module and increase the flexibility of the MIF module in processing features of different scales, so that the model can capture feature information of different scales, this embodiment introduces an integrated attention mechanism module FDT. It captures three types of detailed features at the same time: local, regional and global. Its core is the combination of local self-attention RSA and global self-attention GSA. In order to break the boundary effect between sub-windows, RSA introduces a sliding window strategy, that is, in the process of adjacent layers or iterations, the window is shifted by a certain offset to realize cross-window information flow and feature interaction. This mechanism takes into account both the fine feature aggregation of the local neighborhood and the expansion of the receptive field through orderly window switching, enabling the module to effectively capture local high-frequency information. Under the premise of keeping the computational complexity under control, the model's sensitivity to microstructure is significantly enhanced.

[0044] In addition, this embodiment also improves the upsampling module by using a DUB module, the structure of which is as follows: Figure 9 As shown, this module first upsamples the feature map's dimensions using Dysample, then enhances the upsampled feature map using a 3×3 depthwise convolution (DwConv). To address the issue of information independence between channels in depthwise convolution, a shuffle operation is added. This operation rearranges the channels of the feature map according to certain rules, allowing information from different channels to cross-reference and merge, thereby improving the network's information flow and expressive power. Finally, a 1×1 convolution is used to reduce the number of channels to match the next stage.

[0045] The core module involved in the DUB module is the upsampling operator Dysample, whose working principle is as follows: Figure 10 The first block diagram in the document shows that Dysample generates positional offsets based on input feature content, replacing the traditional dynamic convolution kernel. The specific implementation of the sampling point generator is as follows: Figure 10 As shown in the second block diagram, the low-resolution features are first input, and an initial offset is generated by projecting them using a 1×1 linear projection. Then the offset is reshaped and added to the standard mesh. Above. Then, use the gridsample operation to sort according to the sample set. Input feature map from low resolution Medium resampling yields high-resolution features Subsequently, a bilinear initialization strategy was introduced. The bilinear initialization operation sets the starting point of each target position to the sub-pixel coordinates corresponding to the bilinear interpolation, which avoids the position overlap problem of traditional nearest-neighbor initialization. The principle is as follows: Figure 11 As shown in (a), (b), and (c) in the figure. Taking four points as an example, the points and regions of different colors in the figure represent the initial sampling positions and offset ranges, respectively. In the bilinear initialization strategy, the initial positions are separated to make them evenly distributed, as shown in Figure (b). If the offset processing is not performed on (b), the offset ranges will usually overlap. Therefore, in Figure (c), the offset range is locally constrained to reduce the overlap. That is, the magnitude of the offset is constrained by the dynamic range factor, which can prevent the sampling points from going out of bounds while retaining the flexibility of local adjustment.

[0046] To demonstrate the detection performance of the method described in this embodiment, all experiments were conducted on the DOTA dataset, maintaining the same experimental environment and parameters for each set of experiments. The DOTA dataset is a large-scale image dataset for target detection in aerial images, widely used as a benchmark dataset in the field of remote sensing image target detection. Its images are acquired from multi-source satellites and aerial sensors, covering large-scale scenes with different regions, seasons, and imaging conditions. The dataset contains 2806 aerial images, including 15 categories such as airplanes, ships, tanks, small vehicles, and helicopters, with each image's pixel size ranging from 800*800 to 4000*4000. Since the frequency of some categories varies in natural scenes, studies typically select the most frequent categories for model training based on the distribution of instance numbers. Notably, small targets account for over 90% of the total instances in the dataset, highlighting the importance of small target detection in this field. To meet the needs of model training and evaluation, the original data was divided into training, validation, and test sets in a 7:2:1 ratio to ensure a balanced distribution of each category in the data partitioning.

[0047] Following the third step described above, the MIFF-YOLO model is constructed, pre-trained weights are loaded and frozen into the backbone network, and the parameters of the feature fusion network and the detection head are initialized. Following the fourth step described above, the MIFF-YOLO network is trained using the training samples prepared in the previous steps. During training, 16 sample images are input in each batch, with an initial learning rate of 0.0005 and a final learning rate adjusted to 0.005. A cosine annealing strategy is used for learning rate scheduling, smoothly adjusting the rate over 300 training epochs according to a half-cosine cycle. This configuration allows for sufficient parameter optimization while maintaining training stability.

[0048] After model training, its detection performance was systematically evaluated using a test set. To verify the performance advantages of our proposed method in road target detection, we conducted a horizontal comparison with several mainstream detection algorithms. All comparison models were trained on the same training set and their performance was tested in a unified testing environment. Evaluation metrics included frame-per-second (FPS) detection time, number of model parameters, computational complexity (FLOPs), and average accuracy (mAP@50) and mAP@50:95) for targets of different scales. Accuracy calculations included the entire prediction and post-processing process. Detailed comparative experimental results are shown in Table 1.

[0049] Table 1 Evaluation metrics for small target detection results of different algorithms

[0050] As shown in Table 1, the experimental results demonstrate that the MIFF-YOLO model proposed in this embodiment significantly improves the recognition accuracy of small targets in remote sensing images while maintaining the original detection efficiency. Its balanced parameter size, computational complexity, and detection speed provide a solid foundation for subsequent practical deployment on edge computing devices, effectively solving the technical challenge of small target detection in remote sensing scenarios. The MIFF-YOLO model is comparable to mainstream detection algorithms in terms of parameter size and computational complexity, and exhibits superior lightweight characteristics compared to some algorithms. Regarding detection accuracy, this method significantly outperforms the comparative algorithms in both mAP@50 and mAP@50:95, while maintaining a real-time detection speed of 120 FPS, meeting the processing efficiency requirements of practical deployment and facilitating subsequent deployment.

[0051] To address the challenge of missed or false detections of small targets in remote sensing images, this embodiment systematically improves upon the YOLO11 framework: First, the C3k2 module in the original network is replaced with the C3k2-EFEM module, which integrates edge features and dynamic convolution, effectively enhancing the preservation of spatial details. Second, a cross-dimensional feature interaction MIF module is designed, achieving deep fusion of multi-scale features through 3D convolution and a full-domain attention mechanism. Simultaneously, a DUB module with a dynamic sampling mechanism is introduced to replace the standard upsampling operation, significantly improving the reconstruction quality of small target details and enhancing the model's adaptability to occlusion and complex lighting conditions.

[0052] The above descriptions are merely preferred embodiments of this application, and the present invention is not limited to the above embodiments. It is understood that other improvements and variations directly derived or conceived by those skilled in the art without departing from the spirit and concept of the present invention should be considered to be included within the protection scope of the present invention.

Claims

1. A method for detecting small targets in remote sensing images based on high-dimensional space and multi-scale fusion, characterized in that, The method includes: A dataset is constructed by acquiring a certain number of aerial and satellite remote sensing images of different scenarios, and then divided into training, validation and test sets according to a predetermined ratio. A MIFF-YOLO model for fusing high-dimensional spatial information and multi-scale features is constructed. This model includes a feature extraction network, a feature fusion network, and three decoupled detection heads connected in sequence; wherein: In the feature fusion network, a C3k2-EFEM module is designed using the Sobel detection operator and dynamic convolution ODConv; a MIF module for fusing multidimensional information is designed using the full-domain observer FDT, and the MIF module finally uses 3DMaxPool to compress the computation and maintain the feature representation capability; a dynamic upsampling DUB module is designed using the Dysample operator, and the DUB module further enhances the feature map after Dysample upsampling through depthwise convolution; The training set is input into the MIFF-YOLO model for training. During the training process: The extracted feature information is processed by the feature extraction network. In addition to forward propagation, the specified feature layer information is also passed to the MIF module, where it is effectively fused with the subsequent feature information. At the same time, the feature information is upsampled by the DUB module and then enters the C3k2-EFEM module. The trained MIFF-YOLO model was tested using the test set and validation set to obtain performance metrics and verify the effectiveness of the model. The model that passed the test was then applied to small target detection in remote sensing images.

2. The method for detecting small targets in remote sensing images based on high-dimensional space and multi-scale fusion according to claim 1, characterized in that, In the MIF module, the FDT is used to capture three types of detailed features simultaneously: local, regional, and global. Its core lies in the combination of the regional self-attention mechanism RSA and the global self-attention mechanism GSA.

3. The method for detecting small targets in remote sensing images based on high-dimensional space and multi-scale fusion according to claim 2, characterized in that, In the RSA, the input feature map is divided into non-overlapping sub-windows of uniform size. , where the value of i ranges from 1 to M, and each sub-window covers an N×N spatial region; Within each sub-window, pointwise convolution and depthwise convolution are used to generate the query matrix. Key matrix Sum matrix The dimensions of these matrices are all ; The region self-attention representation of each sub-window feature block is as follows: in, express Activation function This represents a learnable parameter used to control the smoothness of the attention distribution.

4. The method for detecting small targets in remote sensing images based on high-dimensional space and multi-scale fusion according to claim 2, characterized in that, In the GSA, the normalized features of the input layer First, a query matrix is ​​generated through 1×1 pointwise convolution and then 3×3 depthwise convolution. Key matrix Sum matrix The shape was adjusted to have h heads, among which This is the number of channels per head; Attention mapping is expanded along the channel direction, and then vectors are calculated. and The dot product is obtained The global attention graph; the entire process of the global self-attention is represented as follows: in, This represents a learnable parameter.

5. The method for detecting small targets in remote sensing images based on high-dimensional space and multi-scale fusion according to claim 3, characterized in that, The RSA introduces a sliding window strategy, including: During adjacent levels or iterations, the window is shifted by a certain offset to enable cross-window information flow and feature interaction.

6. The method for detecting small targets in remote sensing images based on high-dimensional space and multi-scale fusion according to claim 1, characterized in that, The dynamic upsampling process of Dysample includes: The input feature map is used to generate the offset through a 1×1 convolution. ; The offset Used to adjust the standard sampling grid This yields a new dynamic sampling coordinate set. ; Using the standard sampling grid According to the sampling set Resampling is performed on the original low-resolution input feature map to generate a high-resolution output feature map. The overall process can be represented by the following formula: in, This represents the low-resolution input feature map to be upsampled. It is a 1×1 convolution. It is the sampling grid function.

7. The method for detecting small targets in remote sensing images based on high-dimensional space and multi-scale fusion according to claim 1, characterized in that, In the C3k2-EFEM module, its efficient front-end module EFEM enriches feature information through a dual-branch structure, namely the SobelConv branch and the MaxPool branch; wherein: The SobelConv branch uses the horizontal and vertical convolution kernels of the Sobel operator to explicitly extract edge features in the image. This process involves virtually expanding the feature map, performing convolutions in the X and Y directions respectively, removing the virtual dimension, and finally merging the edge responses in the two directions. The MaxPool branch extracts and retains rich spatial detail information through an additional max-pooling layer.

8. The method for detecting small targets in remote sensing images based on high-dimensional space and multi-scale fusion according to claim 1, characterized in that, The feature extraction network incorporates a C2PSA module, which enhances feature representation capabilities through a position-sensitive attention mechanism; wherein: The C2PSA module employs a multi-head attention structure combined with a feedforward neural network, enabling the model to focus on key information at different spatial locations. Simultaneously, the feedforward neural network maps features to a high-dimensional space and mines complex relationships between features through nonlinear transformations, thereby constructing a more discriminative feature representation.

9. The method for detecting small targets in remote sensing images based on high-dimensional space and multi-scale fusion according to claim 1, characterized in that, In the decoupled detection head structure, the three detection heads use multiple parallel branches to process feature information at different scales. This structure separates the two tasks of target localization and category prediction, allowing them to be learned by independent network modules, and finally the prediction results of the two are fused in the output stage.

10. The method for detecting small targets in remote sensing images based on high-dimensional space and multi-scale fusion according to claim 1, characterized in that, The training process also includes: After co-optimizing the bounding box regression and classification tasks using the DIOU loss function and DFL loss, a task-aligned sample matching strategy is further employed to assign positive and negative samples, including: Calculate the intersection-union ratio (IUU) between each predicted bounding box and its corresponding ground truth bounding box, and multiply it by the classification confidence of the corresponding sample to obtain a task alignment score that reflects the consistency between classification and localization. Based on this, for each true bounding box, the top N predicted samples with the highest alignment scores are selected as positive samples, and the rest are considered as negative samples.

Images